Collaborative-Online-Learning-Enabled Distributionally Robust Motion Control for Multi-Robot Systems

Chao Ning, Han Wang, Longyan Li, Yang ShiarXiv 2025 · DOI · arXiv

Introduction

In real-world multi-robot deployments, navigating environments with unpredictable obstacles—such as humans or other vehicles—remains a significant challenge. These obstacles often exhibit complex, multi-modal motion patterns that are difficult to model accurately in real-time. Furthermore, partial observability due to occlusions often prevents a single robot from gathering enough data to characterize obstacle behavior reliably. Standard motion control approaches frequently struggle with non-Gaussian uncertainties and the computational overhead of centralized data processing.

To address these challenges, this paper introduces COOL-DRMC (COllaborative-Online-Learning-enabled Distributionally Robust Motion Control), a decentralized framework designed to enable a team of robots to collaboratively learn unknown motion patterns and ensure safe navigation through distributionally robust optimization.

Methodology: The COOL-DRMC Framework

The COOL-DRMC framework bridges the gap between collaborative learning and robust control by sharing learning structures rather than raw sensor data.

Collaborative Online Learning (COOL)

At the heart of the system is a non-parametric Bayesian learning approach using a Dirichlet Process Mixture Model (DPMM). Instead of sharing high-bandwidth raw data, robots exchange "learning structures"—specifically, the sufficient statistics (means and covariances) of "clumps" of observed data. This allows the team to automatically identify the number of motion "modes" an obstacle possesses without prior knowledge.

Ambiguity Set Construction and Management

Rather than relying on a single point estimate of the obstacle's distribution, the robots construct an ambiguity set. This set contains all mathematically plausible distributions that align with the observed data, providing a safety buffer against estimation errors.

Overview of the COOL-DRMC framework showing the collaborative learning and ambiguity set management cycle.

As the prediction horizon increases, the number of possible motion combinations can grow exponentially. To maintain computational efficiency, the authors developed a Safe Compression Scheme. This algorithm merges similar components within the ambiguity set while providing theoretical guarantees that the true distribution remains within the compressed set.

Tractable Optimization

To ensure real-time feasibility, the framework reformulates complex probabilistic collision constraints into Semi-Definite Programming (SDP) problems. By utilizing separating hyperplanes, the robust optimization problem becomes solvable in milliseconds, even for multi-robot systems.

Experimental Evaluation

The framework was evaluated through rigorous numerical simulations across three scenarios:

  1. 3D Quadrotors: Navigating around obstacles with multi-modal Gaussian mixture motion.
  2. Nonlinear Vehicles: A 2D intersection scenario with high obstacle uncertainty.
  3. Large-Scale Scalability: Tests involving up to 24 quadrotors to compare decentralized and centralized performance.

Key Results

  • Safety Performance: COOL-DRMC achieved a 97.6% collision-free rate in occluded environments, significantly outperforming individual learning (68.8%) and offline methods (46.4%).
  • Tracking Accuracy: The method demonstrated a 35.7% reduction in tracking error compared to standard Wasserstein-based robust controllers.
  • Scalability: The decentralized approach maintained near-constant computation time as the number of robots increased, whereas centralized methods scaled linearly and became prohibitively slow.

Discussion and Future Work

COOL-DRMC provides a robust, distribution-agnostic safety mechanism that excels in complex environments. However, certain limitations remain. The current model assumes zero-delay communication between robots, which may not hold in high-interference settings. Additionally, while the framework handles translational motion uncertainty well, rotating or pivoting obstacles may require more sophisticated geometric modeling. Future work will also look into finding "robust terminal sets" for more cluttered environments to further enhance long-term tracking guarantees.

Summary updated

BibTeX
@misc{ning2025collaborative,
  author = {Chao Ning and Han Wang and Longyan Li and Yang Shi},
  title = {{Collaborative-Online-Learning-Enabled Distributionally Robust Motion Control for Multi-Robot Systems}},
  year = {2025},
  eprint = {2508.17173},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.17173},
}
Figure